{"id":"10b002bf-5ab7-41d9-aa37-6348b89c9f38","arxiv_id":"2509.03975","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using contrast-enhanced MRI only during training, a multi-task Y-Net improves non-contrast liver vessel segmentation, with the largest gains when annotations are scarce.","lead":"This paper trains a network to segment liver vessels from non-contrast MRI by also teaching it, during training only, to turn those images into contrast-enhanced versions. The extra task improves vessel segmentation, especially when few expert labels are available, and the same recipe helps brain tumor segmentation in a second dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates the auxiliary contrast-enhanced modality from the effect of simply adding unlabeled training data; the reported gains may reflect extra data rather than task compatibility.","rationale":"The reader's weakest_assumption focused on task compatibility and the lack of an ablation isolating NDDR fusion, which is a related but distinct concern. My concern is more specific: the comparison of Y-Net-mix to Y-Net/U-Net confounds the presence of the auxiliary T1wce modality with the presence of 30 additional unlabeled training volumes. This is a confound that directly threatens the paper's central mechanistic claim ('the feature representation benefits from the shared task structure'), not just its statistical robustness. The existing experiments provide no control that uses the same unlabeled non-contrast data without the contrast-enhanced auxiliary task, so the observed gains could be a generic semi-supervised effect. This does not change the verdict because the reader's CONDITIONAL verdict already requires additional evidence; my concern specifies a concrete control that would address it. I recommend keeping the verdict CONDITIONAL (i.e., UNCHANGED) pending such an ablation.","tokens_in":15988,"tokens_out":7104,"duration_ms":64491,"concrete_test":"Train the same Y-Net architecture on the same 12 annotated triplets plus the same 30 unlabeled T1w volumes, but replace the auxiliary T1wce decoder with a decoder that reconstructs the input T1w volume (self-reconstruction). Keep all other settings identical (loss weighting, epochs, augmentation, optimization). If the Dice score on the hold-out test set (and in the 1-annotation setting of Fig. 4a) is statistically indistinguishable from the reported Y-Net-mix results (0.504 hold-out Dice; 0.539 cross-validation with 30 pairs), then the improvement is due to additional unlabeled data rather than the contrast-enhanced auxiliary modality, and the mechanistic claim collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the contrast-enhanced auxiliary modality improves segmentation via shared task structure (abstract; Section 1.4). But the experiments do not control for the fact that the Y-Net-mix is trained with 30 additional unlabeled volumes (T1w/T1wce pairs) compared to the Y-Net and U-Net baselines, which use zero unlabeled data. Any semi-supervised or pretraining approach using the same 30 T1w volumes (e.g., T1w-to-T1w reconstruction) could produce the same gain without the auxiliary contrast-enhanced modality. Section 3.4 states that on the unlabeled pairs 'only the image translation decoder and the shared encoder are trained' (with the segmentation decoder frozen), so this is essentially encoder pretraining. The Y-Net-TL in Section 3.5 is even more explicitly pretraining on the translation task. Without a control that uses unlabeled T1w data alone (without T1wce), the observed improvements (e.g., Dice from 0.506 to 0.539 in Section 3.4) cannot be attributed to the auxiliary modality or to task compatibility; they may simply reflect additional unlabeled training data. The abstract's assertion that 'the feature representation benefits from the shared task structure' is therefore unsupported by the current experimental design.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-task learning framework, Y-Net, for liver vessel segmentation from non-contrast T1w MRI. A shared encoder feeds two decoders: one for vessel segmentation and one for predicting contrast-enhanced T1wce images, with NDDR feature fusion and an uncertainty-weighted loss. Unannotated T1w/T1wce pairs are exploited either by joint training (Y-Net-mix) or by pretraining/fine-tuning (Y-Net-TL). Experiments on liver MRI (12 annotated patients in 3-fold CV plus 3 hold-out cases, with up to 30 unannotated pairs) and on a public brain tumor dataset (8 annotated, 30 auxiliary cases) report Dice improvements, especially with few annotations. The central claim is that an auxiliary contrast-enhanced modality available only during training improves segmentation through shared task structure.","tokens_in":16330,"tokens_out":6045,"duration_ms":57729,"significance":"If the claimed effect is real, the practical value is high: it would allow models to benefit from richer imaging modalities during training while deploying on cheaper, more widely acquired non-contrast sequences, exactly the setting the authors target. The paper has clear strengths: the code is released, the brain-tumor validation uses a public dataset, the vessel-diameter analysis is informative, and the authors explicitly acknowledge several limitations (small cohort, unstable uncertainty weighting). However, as detailed below, the experimental design does not currently isolate the contribution of the auxiliary contrast-enhanced modality from the contribution of simply adding unlabeled data, and the primary non-confounded MTL comparison is not statistically significant. The central attribution to 'shared task structure' therefore needs an additional control experiment.","major_comments":[{"comment":"The headline quantitative claim (Dice 0.506→0.539, p<0.01) is confounded: Y-Net-mix uses 30 additional unlabeled T1w/T1wce volumes, while both U-Net and Y-Net use zero unlabeled volumes. In the unlabeled phase only the shared encoder and translation decoder are trained (§2.2), so the gain may simply reflect encoder pretraining on extra T1w data. There is no control using the same 30 unlabeled T1w volumes with an auxiliary task that does not require T1wce (e.g., T1w reconstruction/inpainting or a self-supervised T1w-only objective). Without such a control, the abstract's statement that 'the feature representation benefits from the shared task structure' is unsupported. The same confound applies to the brain experiment (§3.9). This is the central missing experiment.","section":"§3.4/Table 3 and §2.2"},{"comment":"The only experiment that isolates multi-task learning from additional unlabeled data is the Y-Net vs U-Net comparison on the same 12 annotated patients, which gives Dice 0.506 vs 0.446 (p=0.09) and MSD 4.615 vs 3.929 (p=0.26). This is not statistically significant. The significant p<0.01 gains in Tables 3 and 4 all come from models that also use 5–30 unlabeled pairs—exactly the confounded comparison above. Please report paired per-fold/per-patient statistics, confidence intervals, and specify the test used; as written, the central claim rests on a non-significant primary comparison plus confounded auxiliary-data comparisons.","section":"§3.3/Table 2"},{"comment":"In the Y-Net-mix experiment the uncertainty weights σ_S and σ_T are not learned; they are fixed to values 'learned in the first experiment' because the automated search is unstable under class imbalance. This means the reported Y-Net-mix results do not actually exercise the proposed adaptive loss balancing. The fixed values are selected from a model trained on the same small 12-patient cohort, not an independent validation set, so there is a risk of subtle selection bias. Please disclose the fixed numeric values, justify their transfer to different annotation/pair counts, and report sensitivity to these values (e.g., ±20%) or retrain in a way that makes the uncertainty weights stable.","section":"§3.4 and Eq. (1)"},{"comment":"On the hold-out test set, Y-Net-TL (30 auxiliary pairs) is worse than the U-Net baseline (Dice 0.456 vs 0.474), and Y-Net-mix improves only modestly (0.504). The paper attributes the Y-Net-TL failure to one outlier in the >15 mm group, but Table 7 shows Y-Net-TL is also lower than U-Net in the 10–15 mm group (0.52 vs 0.53) and only marginally better in the small-vessel groups. This contradicts the broad statement in §4 that 'an auxiliary training modality improves segmentation accuracy' when framed as a general cross-domain claim. Please temper the generalization claim or provide a detailed analysis of why Y-Net-TL fails on hold-out data; as written, the evidence does not support the conclusion that auxiliary-data transfer is universally beneficial.","section":"§3.8/Table 6"}],"minor_comments":[{"comment":"'Adams optimizer' should be 'Adam optimizer' (also used inconsistently with Section 3.2).","section":"Table 1"},{"comment":"The mean surface distance for U-Net in the 10–15 mm group is reported as 'inf'. Please state how undefined values (e.g., empty predictions) are handled in the aggregate statistics and whether this affects the reported means.","section":"§3.7/Table 5"},{"comment":"No statistical test is described for any p-value. Specify whether paired t-test, Wilcoxon signed-rank, or another test is used, and whether the unit of analysis is per-fold, per-subject, or per-volume. Also clarify whether the same folds are used across all methods and unlabeled-pair counts.","section":"§3.2–§3.3"},{"comment":"The split into 12 patients for 3-fold cross-validation and 3 hold-out patients is stated only later in the text. Please state this split clearly before the experiments and explain how the 15 annotated triplets are partitioned.","section":"§3.2"},{"comment":"In Fig. 4(a), the x-axis is labeled 'Annotated volumes used for training', but the figure caption says 'cases'. Please be explicit that in this experiment the unlabeled-pair count is fixed at 30 for Y-Net-mix and Y-Net-TL, so the comparison across annotation counts is not confounded by pair count.","section":"§3.6/Fig. 4"},{"comment":"The final paragraph on portal hypertension, HVPG, and clinical outlook is interesting but reads as disconnected from the presented experiments. Consider moving it to a separate 'clinical implications' section or shortening it.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for Computerized Medical Imaging and Graphics and addresses a practically important problem. The main barrier is experimental, not conceptual: the missing control with unlabeled T1w-only auxiliary data is feasible and would directly test the paper's core claim. I would be willing to review a revised version. I also note that the fixed σ_S/σ_T values and the small cohort make the quantitative claims fragile, but these are fixable with additional experiments and clearer statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you work on medical image segmentation with scarce annotations. The setup is genuinely new—using paired contrast-enhanced and non-contrast exams at train time, with the contrast-enhanced arm dropped at inference, via a shared encoder and NDDR fusion—and the paper is honest and reproducible: code is public, they validate on liver vessels and brain tumors, and they explicitly flag their small cohort and unstable loss weighting. That part is solid.\n\nThe soft spot is the missing control for the auxiliary modality. Y-Net-mix is trained with 30 extra unlabeled T1w/T1wce pairs, while the U-Net and Y-Net baselines use zero unlabeled data. So the reported gains (e.g., Dice 0.506 to 0.539) could come from the extra unlabeled images alone, via any semi-supervised objective, and not from the fact that the auxiliary modality is contrast-enhanced. Section 3.4 says the unlabeled pairs only train the translation decoder and shared encoder—essentially pretraining on a synthesis task. Y-Net-TL is explicitly pretraining. There is no control that uses unlabeled T1w data without T1wce (say, a reconstruction task). Without that, the abstract's claim that 'the feature representation benefits from the shared task structure' is not supported. The primary U-Net vs Y-Net comparison is also not significant (p=0.09); the significant gains come from adding the unlabeled pairs.\n\nThis does not destroy the practical usefulness—if using routinely acquired contrast-enhanced pairs improves segmentation, that is useful even if the benefit is mostly from more training data. But the paper needs a control to know what is doing the work. If the authors can run that, the claim becomes much stronger. As is, it is a well-executed empirical study with a load-bearing identifiability issue. The brain validation helps but has the same gap.\n\nThis paper deserves a serious referee: the question is important, the two-domain evidence is valuable, and the missing control is fixable. For a reading group, it would generate good discussion about experimental design in semi-supervised multi-task learning. I would not cite it as evidence for the auxiliary-modality benefit until the control is in place.","headline":"Well-executed medical-imaging study with a genuinely new setup, but the central claim that the auxiliary contrast-enhanced modality drives the gains is not isolated from simply adding unlabeled training data.","tokens_in":16783,"tokens_out":4089,"would_cite":false,"duration_ms":38482,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An auxiliary contrast-enhanced MRI sequence available only during training still improves vessel segmentation from non-contrast MRI, and the gain is largest when expert annotations are scarce.","keywords":["multi-task learning","vessel segmentation","liver MRI","auxiliary contrast-enhanced MRI","semi-supervised segmentation","Y-Net","brain tumor segmentation","deep learning segmentation"],"falsifier":"Train the same Y-Net with identical unannotated pairs but replace the auxiliary T1wce target with a task known to be irrelevant to vessels—for example, translating the image to a scrambled or unrelated output while keeping the same loss and data size. If the Dice gain over the U-Net persists, the benefit comes from the extra training signal rather than from the informative contrast-enhanced modality; if the gain disappears, task compatibility is necessary.","tokens_in":15910,"feed_emoji":"🩻","tokens_out":7073,"duration_ms":61922,"temperature":0.7,"pith_summary":"The paper tries to establish that a richer imaging modality can be used during model training and then discarded at deployment, while still improving predictions made from the routine modality. In this test case, a multi-task network segments liver vessels from non-contrast T1w MRI by simultaneously learning to translate T1w into contrast-enhanced T1wce MRI during training; the T1wce images are never needed at test time. With identical annotations, this raises the Dice overlap score from 0.446 for a U-Net to 0.506 for the multi-task Y-Net, and adding 30 unannotated T1w/T1wce pairs pushes it to 0.539 (p<0.01). The advantage is strongest when only one or a few annotated volumes exist—for a single annotation, Dice rises from 0.242 to 0.372. The same recipe improves brain tumor segmentation from T2w MRI when T1wce is the training-only auxiliary, suggesting the strategy is not specific to one organ or sequence.","feed_headline":"Training-only contrast MRI lifts vessel Dice from 0.45 to 0.54","feed_subtitle":"Multi-task Y-Net uses contrast-enhanced images it never sees at deployment, and gains most when annotations are scarce.","key_machinery":"The Y-Net architecture: a 3D U-Net with one shared encoder and two task-specific decoders, one for vessel labels (T_S) and one for translating the input T1w image into a contrast-enhanced T1wce image (T_T). The two decoders exchange features at matching spatial resolutions through Neural Discriminative Dimensionality Reduction (NDDR) layers—1x1x1 convolutions that learn how to fuse concatenated decoder features. A homoscedastic-uncertainty loss with learned weights balances segmentation cross-entropy and translation mean-squared error. During training, these components let the model use both annotated triplets and unannotated image pairs; at inference, only the non-contrast image enters and","core_discovery":"The central claim is that auxiliary contrast-enhanced MRI, available only during training, improves segmentation of vessels in non-contrast MRI, and does so most when annotated examples are scarce. A Y-Net with a shared encoder and two decoders—one predicting vessel labels, one predicting the contrast-enhanced volume—learns both tasks together. Features flow between the two decoders through layerwise fusion layers, so the translation task acts as a training-time teacher for the segmentation task. Unannotated pairs of non-contrast and contrast-enhanced images can be added to the training set, and they make the segmentation better even though no vessel labels come with them. On the liver datas","pith_inferences":["The underlying principle should generalize beyond contrast/non-contrast MRI pairs: any auxiliary imaging modality that reveals structure relevant to the target label—such as a later CT phase, a higher-resolution sequence, or an alternative contrast mechanism—could be used the same way, provided the two tasks are compatible.","A direct testable extension would be to compare this multi-task approach against self-training or pseudo-labeling on the same unannotated pairs; if the Y-Net wins, the shared-task structure, not merely the extra images, is carrying the gain.","The paper does not isolate how much of the benefit comes from the NDDR fusion layers versus simply pretraining the encoder on translation; an ablation that removes the cross-decoder connections would separate those two contributions.","Clinical deployment of the idea would require checking that the model's improved recall does not come with silently worse precision for the structures a radiologist cares about; the reported precision trade-off makes this worth measuring per vessel diameter."],"forward_implications":["If the claim is right, contrast-enhanced sequences can be treated as a training-only resource: hospitals can keep routine non-contrast protocols at deployment and still get better vessel segmentation than a model trained on non-contrast data alone.","The method directly addresses the low-annotation regime: with a single annotated liver volume, the multi-task model with 30 unannotated pairs reaches Dice 0.372 versus 0.242 for the U-Net, and the gap shrinks as annotations increase.","Unannotated paired images without labels provide most of the benefit: Y-Net-mix adding 30 such pairs outperforms the plain Y-Net and matches or beats pretraining-plus-fine-tuning (Y-Net-TL) in cross-validation, while on the hold-out set only Y-Net-mix beats the U-Net.","The benefit transfers across domains: the same architecture and training recipe improve brain tumor segmentation from T2w MRI using T1wce as the auxiliary modality, indicating the approach is not tuned to liver vessels.","Improvements are not uniform across vessel sizes: the largest Dice gains occur for vessels above 10 mm in diameter, while small vessels (0–5 mm) remain difficult and still need more annotated examples."],"supporting_citations":[{"why":"Supplies the 3D U-Net backbone on which all segmentation models are built.","marker":"[27]"},{"why":"Y-Net one-to-two architecture; supplies the shared-encoder/two-decoder structure used for segmentation plus translation.","marker":"[41]"},{"why":"NDDR layers; the layerwise feature-fusion mechanism that lets the two decoders share learned features.","marker":"[42]"},{"why":"Uncertainty-based loss weighting; balances the segmentation and translation losses with learned parameters.","marker":"[36]"},{"why":"Medical Segmentation Decathlon brain tumor dataset; provides the cross-domain validation with T2w input and T1wce auxiliary data.","marker":"[47]"},{"why":"Baseline 3D U-Net implementation used for the single-task comparison.","marker":"[50]"}],"fun_headline_variants":["Training-only contrast MRI lifts vessel Dice","Scarce vessel labels? Contrast MRI at training helps","Multi-task Y-Net gains from contrast data it never sees","Unlabeled contrast pairs boost vessel segmentation","Auxiliary contrast imaging only during training improves vessels"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that translating a non-contrast image into its contrast-enhanced version teaches features that also help find vessels; if the translation and segmentation tasks are not actually compatible, the gains rest on extra data alone and the paper's explanation for them collapses.","fun_headline_variants_meta":{"raw":{"variants":["Training-only contrast MRI lifts vessel Dice","Scarce vessel labels? Contrast MRI at training helps","Multi-task Y-Net gains from contrast data it never sees","Unlabeled contrast pairs boost vessel segmentation","Auxiliary contrast imaging only during training improves vessels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000345,"raw_usage":{"total_tokens":1717,"prompt_tokens":716,"completion_tokens":1001,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":928}},"tokens_in":460,"tokens_out":1001,"duration_ms":9340,"temperature":1.0,"reasoning_tokens":928,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:29:11.354672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same Y-Net with identical unannotated pairs but replace the auxiliary T1wce target with a task known to be irrelevant to vessels—for example, translating the image to a scrambled or unrelated output while keeping the same loss and data size. If the Dice gain over the U-Net persists, the benefit comes from the extra training signal rather than from the informative contrast-enhanced modality; if the gain disappears, task compatibility is necessary.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 3D U-Net backbone on which all segmentation models are built."},{"cited_title":"Multi-Task Learning for Dense Prediction Tasks: A Survey","cited_arxiv_id":"2004.13379","evidence_quote":"Y-Net one-to-two architecture; supplies the shared-encoder/two-decoder structure used for segmentation plus translation."},{"cited_title":"Multi- task learning using uncertainty to weigh losses for scene geometry and semantics,","cited_arxiv_id":null,"evidence_quote":"NDDR layers; the layerwise feature-fusion mechanism that lets the two decoders share learned features."},{"cited_title":"Which Tasks Should Be Learned Together in Multi-task Learning?","cited_arxiv_id":"1905.07553","evidence_quote":"Uncertainty-based loss weighting; balances the segmentation and translation losses with learned parameters."},{"cited_title":"Y-net: a one-to-two deep learning framework for digital holographic reconstruction,","cited_arxiv_id":null,"evidence_quote":"Medical Segmentation Decathlon brain tumor dataset; provides the cross-domain validation with T2w input and T1wce auxiliary data."},{"cited_title":"Building skeleton models via 3-d medial sur- face axis thinning algorithms,","cited_arxiv_id":null,"evidence_quote":"Baseline 3D U-Net implementation used for the single-task comparison."}],"review_version":1}